benchmarks: one stage generator, make_2611_stages.py; delete the 2026-10 generators (#430) - #431

make_2611_stages.py is now the only stage generator. It was a profile of make_2610_stages.py, itself a profile of make_october_stages.py, and rewrote the October template with _sub text anchors. The template (HEAD) now takes the instrument, REPS and the banner as parameters (stage_params), the stage table holds only the lanes the 2026-11 subset stages (graph, cross-model, documents, dense, sparse, time series, lifecycle, plus the restricted-arm stages derived from runner.ARM_RUNS), and nothing is rewritten after the fact. DECISIONS #196: "october was dev".

Proof that no stage moved, apart from two intended lines: before editing I recorded sha256 of every draft three generations write with the old chain (comparators at the default order and 3 reps, comparators in paper order at 5 reps, ArcadeDB alone on a release for every lane with the second list). The first commit's drafts, stdout and --check output were identical byte for byte; the second commit then changes exactly two things in every lane stage (the pair-install script is unchanged): the banner no longer names the deleted generators, and the grep -q 'def instrument_at_least' bench_common.py check is gone (it was asked for by #429's owner; bench_common keeps the function). The diff of one draft (qS03.sh, old to new):

4,5c4,5
< # GENERATED by make_2611_stages.py (a profile of make_2610_stages.py and make_october_stages.py) -- edit
< # those files, not this one, and
---
> # GENERATED by make_2611_stages.py -- edit
> # that file, not this one, and
35d34
< grep -q 'def instrument_at_least' bench_common.py || { say "$ID ABORT: no BENCH_INSTRUMENT selector in this tree"; exit 1; }

stage_drafts_2611.sha256 holds the 32 hashes of the new drafts and test_the_drafts_are_the_ones_on_file compares them on every run (regenerate with python3 test_stage_generator_2611.py --rewrite after reviewing a deliberate change).

Deleted: make_october_stages.py, make_2610_stages.py, pass2_roster.py and the tests that only exercised 26.10.x passes (test_pass2_roster.py, test_stage_pass_tiers.py, test_stage_rep_pass.py, test_stage_top_up.py, test_stage_tiers.py, 82 tests), oct_repin_smoke.sh, oct_skeleton_arcadedb.sh, structure_check.check_stage_chain (it held CAMPAIGN.md's October chain against the deleted table), the october_stages/ ignore entry, the host-side Python-cost, e4 and restart stage templates that the 2026-11 generator never emitted, and the CAMPAIGN.md rows and paragraphs that only describe the 2026-10 passes (rows 45, 46, 53, the 26.10.1 measurement paragraph, the October chain drawing). check_coverage stays, now for the lanes this table stages, and --check runs it. Re-pointed: cap_rule.known_backends (reads runner.LANES), test_cap_rule.py, test_jpype_pin.py, test_cell_cost_backends.py, test_jvm_defaults_arm.py, test_stage_generator_after.py, and the comments in bench_common.py, Dockerfile.bench, build_images.sh, queue_lint.py; the docs lines in CAMPAIGN.md, PROTOCOL.md, PUBLISHING.md, READING-RESULTS.md and QUANTIZATION.md.

git grep for the deleted generators' names is now empty.

--arcadedb-release X.Y.Z already emits an ArcadeDB-only stage set on a new release (the pair install, then every ArcadeDB arm with the exactly-once proof), which is what #326 asked of the old generator, so this closes it.

Tests (benchmarks/experiments, nice -n 10 taskset -c 12-15): before 1007 passed, 1 failed, 24 skipped; after 918 passed, 1 failed, 24 skipped. The 89 fewer are the 82 deleted tests, 2 from test_stage_generator_after.py, 4 host-stage jpype tests, 2 fewer cap-rule modes, and 1 added (the sha test). The one failure, test_dense_ef_sweep.py::test_export_adds_the_matched_columns_only_for_2026_11, fails identically on origin/main in this worktree (the worktree lacks the untracked bench data that main's checkout has, which I copied in read-only for the run); it is not touched here. The same suite on this branch merged into #429's branch (instrument-2611-only): 850 passed, same 1 failure, 24 skipped; test_check_refuses_another_instrument_in_the_environment now accepts the refusal from either layer (the generator's own check or bench_common at import). None of my tests read untracked results/. No Maven, Docker, benchmark cells or timed runs. pre-commit passes on every changed file.

Closes #430
Closes #326

🤖 Generated with Claude Code